We performed a search of the databases China National Knowledge Infrastructure (CNKI), Wanfang Database, Cqvip Database (VIP), SinoMed, PubMed, EMBASE, Cochrane Library, Web of Science databases without language restrictions from their inception to April 27, 2022. Taking PubMed as an example, the Medical Subject Headings (MeSH) terms and keywords used in the search were as follows: ((("Pneumonia, Mycoplasma"[Mesh]) OR ((((((((((((((((Refractory Mycoplasma pneumoniae pneumonia[Title/Abstract]) OR (Refractory Pneumonia, Primary Atypical[Title/Abstract])) OR (Refractory Atypical Pneumonia, Primary[Title/Abstract])) OR (Refractory Atypical Pneumonias, Primary[Title/Abstract])) OR (Refractory Pneumonias, Primary Atypical[Title/Abstract])) OR (Refractory Primary Atypical Pneumonia[Title/Abstract])) OR (Refractory Primary Atypical Pneumonias[Title/Abstract])) OR (Refractory Mycoplasma Pneumonia[Title/Abstract])) OR (Refractory Mycoplasma Pneumonias[Title/Abstract])) OR (Refractory Pneumonias, Mycoplasma[Title/Abstract])) OR (Refractory Mycoplasma ovipneumoniae Infection[Title/Abstract])) OR (Refractory Mycoplasma ovipneumoniae Infections[Title/Abstract])) OR (Refractory Mycoplasma pneumoniae Infection[Title/Abstract])) OR (Refractory Mycoplasma pneumoniae Infections[Title/Abstract])) OR (Refractory Mycoplasma dispar Infection[Title/Abstract])) OR (Refractory Mycoplasma dispar Infections[Title/Abstract]))) AND (("L-Lactate Dehydrogenase"[Mesh]) OR (((((LDH[Title/Abstract]) OR (Dehydrogenase, L-Lactate[Title/Abstract])) OR (L Lactate Dehydrogenase[Title/Abstract])) OR (Lactate Dehydrogenase[Title/Abstract])) OR (Dehydrogenase, Lactate[Title/Abstract])))) AND (sensitiv*[Title/Abstract] OR sensitivity and specificity[MeSH Terms] OR (predictive[Title/Abstract] AND value*[Title/Abstract]) OR predictive value of tests[MeSH Terms] OR accuracy*[Title/Abstract]). The references of the included studies and existing systematic reviews were hand-searched to find additional relevant articles. The full search strategy is included in Supplementary Appendix S1.

① Patientor population: Children with Mycoplasma pneumoniae pneumonia; Intervention: Serum LDH testing; Comparison: RMPP compliant with the “gold standard”; Outcome: Final diagnosis of RMPP. ② Studies evaluating LDH as a predictor of RMPP in children. ③ The serum LDH was determined before the diagnosis was made and the corresponding threshold was determined to further evaluate the actual accuracy of LDH for predicting RMPP after confirming or excluding RMPP according to the “gold standard”. ④ The results of the literature must provide sufficient data to construct a 2 × 2 contingency table to accurately calculate the specific values of true positive (TP), false positive (FP), false negative (FN), and true negative (TN) cases when LDH predicted RMPP in this study. ⑤ The type of study design is prospective or retrospective study. ⑥ The study also needs to clearly point out the reference standard of RMPP.

① conference abstracts or study protocols. ② duplicate publications. ③ studies with incomplete data or no relevant outcome.

Two reviewers (Shumin WANG and Xuejun LI) independently screened the literature. Initially, duplicate and irrelevant publications were excluded based on their title and abstract. Later, each of them independently read the full text of each potentially eligible article and finally identified all studies. In case of disagreement, discussions were conducted with a third investigator (Yixing ZHANG) until consensus was reached.

Two reviewers (Shumin WANG and Xuejun LI) independently extracted the followingdata using predesigned forms according to the guideline fordata extraction for systematic reviews and meta-analysis (12), including the following information: first author, year of publication, nation, study type, sample size, gender ratio, average age of participants, “gold standard” for diagnosing RMPP, TP, FP, FN, and TN cases, thresholds of LDH. If the data in the literature is incomplete or cannot be directly extracted, the corresponding author will be contacted by email for consultation, and the literature that has not received an affirmative reply from the author will be excluded.

The Quality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2) was used to assess the quality of the included studies (13). The tool evaluates 4 domains: patient selection, index test, reference standard, and flow and timing. For each domain, the risk of bias is analyzed using different signaling questions. Beyond the risk of bias, the tool also evaluates the applicability of each included study to the research question for every domain.

Meta-analysis was performed using The Stata software (version 14.0) and Meta-disc 1.4 software. Pooled sensitivity, specificity, positive likelihood ratio (PLR), negative likelihood ratio (NLR), and diagnostic odds ratio (DOR) were calculated using a bivariate mixed-effects model, and a comprehensive receiver operating characteristic curve (SROC) was drawn. Threshold effect was assessed by Spearman's correlation coefficient between the log value of sensitivity and the log value of specificity, when P < 0.05, it was considered to have a threshold effect. Heterogeneity was assessed using the I² value, when I² > 50%, indicating that the heterogeneity between studies was large. We performed subgroup and sensitivity analyses to verify the robustness of the overall results and explore the sources of heterogeneity. Subgroup analysis was performed according to factors such as different study sources, study type, mean age, gold standard, and threshold. Publication bias was assessed with Deeks' funnel plot.

We acquired a total of 805 articles from the initial search of which 385 duplicate articles were excluded. Furthermore, 361 unrelated articles were excluded after the title and abstract screening and then 59 articles after full-text reading, Among of them, 1 articles could not obtain the full text, 27 articles could not extract relevant data, 2 articles have incorrect data. Finally, 29 studies were included in this review. The search selection process is shown in Figure 1.

A total of 29 articles were included, including 702,2 subjects, of which 227,9 children were diagnosed RMPP and 474,3 children were diagnosed general Mycoplasma pneumoniae pneumonia (GMPP). 2 studies (14, 15) used diagnostic criterion 1 as the “gold standard” for diagnosing RMPP, and the rest of the studies used diagnostic criterion 2 as the “gold standard” for diagnosing RMPP. The included studies were mainly from Asian countries, of which 1 study (16) was from Japan, 3 studies (14, 15, 17) were from South Korea, and the rest were from China. The sample size of each group ranged from 20 to 703 cases. 2 cohort studies (7, 18) and the rest were case-control studies. The basic characteristics of the included literature are shown in Table 1.

The quality of the included literature was evaluated based on the QUADAS-2 model, and the results are shown in Figure 2. 2 studies (7, 18) were clearly indicated as cohort studies with low risk of bias in patient selection, and the remaining studies were case-control studies, so there may be high risk of bias in patient selection. 2 studies (14, 17) were defined as uncertain risk in index test because they did not clearly state the threshold, and the remaining studies were defined as high risk of bias in index test because their threshold was not determined in advance. All studies were defined as low risk of bias with respect to the reference standard because each study clearly indicated the gold standard for diagnosis and the interpretation of the results was performed without knowledge of the results of the trial to be evaluated. 7 studies (14, 17, 20, 28, 36–38) did not specify the timing of LDH detection, so the flow and timing part was defined as uncertain risk, and the remaining studies were at low risk of bias. Overall, the quality of the included literature can be rated as moderate to high, and the results of all studies are generally credible.

The data was imported into Meta DiSc1.4 software for analysis, and the Spearman correlation coefficient between the logarithm of sensitivity and (1-specificity) logarithm was −0.324 (P = 0.086 > 0.05), which was not significant, meaning that There was no threshold effect in this study. Further, by drawing a symmetrical SROC curve, there is no “shoulder-arm shape”, which further shows that this study has no threshold effect.

The Cochran-Q test of the DOR showed that Cochran-Q = 312.50, P < 0.001, which means that there is heterogeneity caused by non-threshold effect in this study. Further, the I² of sensitivity, specificity and PLR, NLR, and DOR were all greater than 50%. Therefore, the random effects model was used to combine the above five effect sizes.

Pooled sensitivity was 0.75 (95% CI = 0.73–0.76), pooled specificity was 0.73 (95% CI = 0.72–0.74), pooled PLR was 3.61 (95% CI = 2.86–4.56), pooled NLR was 0.30 (95% CI = 0.23–0.39), pooled AUC = 0.85(95% CI = 0.82–0.88), Q index = 0.7803, and the combined DOR was 13.04 (95% CI = 8.24–20.63). Results shown in Figure 3.

Stata 14.0 was used to perform sensitivity analysis on the data of this study, and the results are shown in Figure 4. It can be clearly seen from the figure that there are 2 original studies (7, 20) with strong sensitivity, and other original studies will not cause the sensitivity of the calculation results. Therefore, the heterogeneity test was performed again after removing the above 2 studies. The Cochran-Q test of the DOR showed that Cochran-Q = 216.51, P < 0.001, suggesting that there is still heterogeneity. Pooled sensitivity was 0.80 (95% CI = 0.78–0.81), pooled specificity was 0.73 (95% CI = 0.71–0.74), pooled PLR was 3.86 (95% CI = 3.05–4.88), pooled NLR was 0.27 (95% CI = 0.22–0.34), pooled AUC = 0.86 (95% CI = 0.83–0.89), Q index = 0.7950, and the combined DOR was 15.36 (95% CI = 9.94–23.72). Similar to the results before exclusion. Therefore, the results of this study are relatively stable.

After the heterogeneity test of the results of the 29 studies, it was suggested that there was direct non-threshold heterogeneity among the studies. Therefore, we further performed subgroup analyses according to different study sources, study type, mean age, gold standard, and threshold. The results are shown in Table 2.

The Deeks' funnel plot showed that the included studies were approximately symmetrically distributed (P = 0.36), indicating that there was no significant publication bias. The results are shown in Figure 5.